CosmoGNN

Estimation of cosmological parameter $\Omega_m$ from Quijote simulations using Graph Neural Networks.

Data

The data used in this work can be retrieved from globus following the path:

Halos/FoF/latin_hypercube/

selecting for each simulation number, the folder with redshift $z = 1$:

groups_002

Requisites

The code runs on GPU. The one used for this work is a Tesla T4.

Libraries:

• numpy
• pytorch
• pytorch-geometric
• matplotlib
• scipy
• sklearn
• optuna

Scripts

• clean_main.py: main driver to train and test the network

• clean_hyperparameters.py: definition of the hyperparameters employed by the networks

• clean_gridsearch.py: optimize the hyperparameters using optuna

• visualize_graphs.py: display graphs of DM halos from the simulations

The folder Source contains:

• constants.py: basic constants and initialization

• load_data.py: routines to load data from simulation files

• plotting.py: functions for displaying the results from the training and test

• metalayer.py: definition of the Graph Neural Network architecture

• training.py: routines for training and testing the network

Authors and Acknowledgments

Authors

• Lorenzo Cavezza - Lorycav
• Giulia Doda - giuliadoda
• Giacomo Longaroni - GiacomoLongaroni
• Laura Ravagnani - LauraRavagnani

Acknowledgments

This work is based on:

Reference Papers

[1] Villanueva-Domingo, Pablo, and Francisco Villaescusa-Navarro. "Learning cosmology and clustering with cosmic graphs." The Astrophysical Journal 937.2 (2022): 115.

[2] Makinen, T. Lucas, et al. "The cosmic graph: Optimal information extraction from large-scale structure using catalogues." arXiv preprint arXiv:2207.05202 (2022).

Original Code

PabloVD, (2023). CosmoGraphNet: "Graph Neural Networks to predict the cosmological parameters or the galaxy power spectrum from galaxy catalogs". GitHub